How rauaan/rag-agent works
This repository implements a question-answering assistant for Kazakhstan’s Sovereign Wealth Fund law, with optional live web research. The workflow routes a user query, gathers local and/or web documents, splits and reranks them in a temporary vector store, then generates a source-grounded answer. A separate indexing entry point is present, but its implementation was not excerpted; its connections are therefore omitted. The graph follows the shown calls and documented prompt behavior.
Main entry points: agent.py Answer workflow · router.py Tool router · local_search.py Local search · web_search.py Web search
Components 18
Query orchestration
- Answer workflow
agent.py - Tool router
router.py
Document retrieval
- Local search
tools/local_search.py - Web search
tools/web_search.py - Web search service
- Page converter
- Tool interface
tools/base.py - Persistent store
rag/vectorstore.py - Indexing entry point
index.py
RAG processing
- Document splitter
rag/splitter.py - Temporary store
rag/vectorstore.py - Context builder
utils.py
Models and prompts
- Embedding model
rag/embeddings.py - Language model
llm.py - Routing prompt
prompts/router_prompt.py - Answer prompt
prompts/system_prompt.py - Runtime settings
config.py
Other components
- Question asker
Connections 21
- Question asker → Answer workflow: asks
agent.py - Answer workflow → Tool router: routes query
agent.py - Tool router → Routing prompt: formats query
router.py - Tool router → Language model: classifies query
router.py - Answer workflow → Local search: requests local search
agent.py - Answer workflow → Web search: requests web search
agent.py - Local search → Persistent store: retrieves documents
tools/local_search.py - Web search → Web search service: searches web
tools/web_search.py - Web search → Page converter: converts pages
tools/web_search.py - Answer workflow → Document splitter: splits documents
agent.py - Answer workflow → Temporary store: indexes chunks
agent.py - Answer workflow → Temporary store: retrieves matches
agent.py - Temporary store → Embedding model: embeds documents
rag/vectorstore.py - Persistent store → Embedding model: embeds documents
rag/vectorstore.py - Answer workflow → Context builder: builds context
agent.py - Answer workflow → Answer prompt: formats answer request
agent.py - Answer workflow → Language model: generates answer
agent.py - Language model → Runtime settings: reads model setting
llm.py - Tool router → Tool interface
- Local search → Tool interface
tools/local_search.py - Answer workflow → Question asker: returns answer
agent.py